Data Driven Prognostics for Failure of Power Semiconductor Packages

Power chips such as Metal Oxide Field Effect Transistors (MOSFETs) are widely used and can be found in many electronics and electrical products. The ability to predict the degradation of such power electronic devices can minimise the risk of their failure during operation and support maintenance planning operations. In this study, a data driven prognostics approach using system identification and machine learning modelling technique is developed and used to predict the time-to-failure of MOSFET TO-220 packages associated with delamination failure mode of the die attachment. Run-to-failure data under thermal overstress loading conditions for power chip devices, available from the NASA Prognostics Centre data repository, is used to develop a data-driven prognostic model that can be used to predict the time-to-failure (TtF) of power MOSFETs under accelerated test loads. An increment in ON-state resistance of the MOSFET is used as precursor for device failure through die-attach degradation. Results from this research show that when monitored data from a damage indicator for a particular failure mode of an electronic package changes dynamically, data-driven modelling using engineering control techniques such as State-Space representation is capable of producing reliable, multi-step ahead predictions for the time-to-failure of the device.